Research map / Video / Video-language understanding
AVATAR: Reinforcement Learning to See, Hear, and Reason Over Video
Yogesh Kulkarni, Pooyan Fazli
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add TAS (U-shaped advantage weighting) vs uniform weights (ablation) | AVATAR on Qwen2.5-Omni · OmniBench | accuracy: 47.8 -> 49.1 | Improved |
| add TAS vs uniform weights (ablation) | AVATAR on Qwen2.5-Omni · TOMATO | accuracy: 29.3 -> 30.8 | Improved |
| use TAS instead of inverse parabolic weights (ablation) | AVATAR on Qwen2.5-Omni · OmniBench | accuracy: 46.5 -> 49.1 | Improved |
| use TAS instead of linear decay weights (ablation) | AVATAR on Qwen2.5-Omni · Video-Holmes | accuracy: 44.3 -> 45.1 | Improved |
| use TAS instead of linear incline weights (ablation) | AVATAR on Qwen2.5-Omni · MMVU | accuracy: 65.0 -> 65.6 | Improved |
| add replay buffer to GRPO (TAS-free) (ablation) | GRPO on Qwen2.5-Omni · OmniBench | accuracy: 45.4 -> 47.8 | Improved |
| add TAS to on-policy GRPO (ablation) | GRPO on Qwen2.5-Omni · VSI-Bench | accuracy: 26.5 vs 26.1 for buffer | Improved |
| add both replay buffer and TAS to GRPO (ablation) | GRPO on Qwen2.5-Omni · Video-MMMU | accuracy: 48.1 -> 49.4 | Improved |
| add off-policy replay buffer (ablation) | AVATAR on Qwen2.5-Omni · WorldSense Hallucination | accuracy: 40.1 -> 51.2 | Improved |
| add TAS (ablation) | AVATAR on Qwen2.5-Omni · WorldSense Audio Change | accuracy: 35.8 -> 37.3 | Improved |
| add SFT cold start (ablation) | AVATAR on Qwen2.5-Omni · OmniBench | accuracy: 43.8 -> 49.1 | Improved |
| add SFT cold start (ablation) | GRPO on Qwen2.5-Omni · OmniBench | accuracy: 41.2 -> 46.0 | Improved |
Closest papers by meaning
- Omni-R1: Reinforcement Learning for Omnimodal Reasoning via Two-System Collaboration (NeurIPS 2025 · similarity 0.89)
- Find, Fix, Reason: Context Repair for Video Reasoning (ICML 2026 · similarity 0.88)
- ReAgent-V: A Reward-Driven Multi-Agent Framework for Video Understanding (NeurIPS 2025 · similarity 0.87)
- Perception-Aware Policy Optimization for Multimodal Reasoning (ICLR 2026 · similarity 0.87)
- PyVision-RL: Forging Open Agentic Vision Models via RL (ICML 2026 · similarity 0.87)
- Incentivizing Versatile Video Reasoning in MLLMs via Data-Efficient Reinforcement Learning (CVPR 2026 · similarity 0.87)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.
